Thu, 22 Jan 2015 13:21:57 +0100
Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6
michael@0 | 1 | //////////////////////////////////////////////////////////////////////////////// |
michael@0 | 2 | /// |
michael@0 | 3 | /// General FIR digital filter routines with MMX optimization. |
michael@0 | 4 | /// |
michael@0 | 5 | /// Note : MMX optimized functions reside in a separate, platform-specific file, |
michael@0 | 6 | /// e.g. 'mmx_win.cpp' or 'mmx_gcc.cpp' |
michael@0 | 7 | /// |
michael@0 | 8 | /// Author : Copyright (c) Olli Parviainen |
michael@0 | 9 | /// Author e-mail : oparviai 'at' iki.fi |
michael@0 | 10 | /// SoundTouch WWW: http://www.surina.net/soundtouch |
michael@0 | 11 | /// |
michael@0 | 12 | //////////////////////////////////////////////////////////////////////////////// |
michael@0 | 13 | // |
michael@0 | 14 | // Last changed : $Date: 2013-06-12 10:24:44 -0500 (Wed, 12 Jun 2013) $ |
michael@0 | 15 | // File revision : $Revision: 4 $ |
michael@0 | 16 | // |
michael@0 | 17 | // $Id: FIRFilter.cpp 171 2013-06-12 15:24:44Z oparviai $ |
michael@0 | 18 | // |
michael@0 | 19 | //////////////////////////////////////////////////////////////////////////////// |
michael@0 | 20 | // |
michael@0 | 21 | // License : |
michael@0 | 22 | // |
michael@0 | 23 | // SoundTouch audio processing library |
michael@0 | 24 | // Copyright (c) Olli Parviainen |
michael@0 | 25 | // |
michael@0 | 26 | // This library is free software; you can redistribute it and/or |
michael@0 | 27 | // modify it under the terms of the GNU Lesser General Public |
michael@0 | 28 | // License as published by the Free Software Foundation; either |
michael@0 | 29 | // version 2.1 of the License, or (at your option) any later version. |
michael@0 | 30 | // |
michael@0 | 31 | // This library is distributed in the hope that it will be useful, |
michael@0 | 32 | // but WITHOUT ANY WARRANTY; without even the implied warranty of |
michael@0 | 33 | // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
michael@0 | 34 | // Lesser General Public License for more details. |
michael@0 | 35 | // |
michael@0 | 36 | // You should have received a copy of the GNU Lesser General Public |
michael@0 | 37 | // License along with this library; if not, write to the Free Software |
michael@0 | 38 | // Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
michael@0 | 39 | // |
michael@0 | 40 | //////////////////////////////////////////////////////////////////////////////// |
michael@0 | 41 | |
michael@0 | 42 | #include <memory.h> |
michael@0 | 43 | #include <assert.h> |
michael@0 | 44 | #include <math.h> |
michael@0 | 45 | #include <stdlib.h> |
michael@0 | 46 | #include "FIRFilter.h" |
michael@0 | 47 | #include "cpu_detect.h" |
michael@0 | 48 | |
michael@0 | 49 | #ifdef _MSC_VER |
michael@0 | 50 | #include <malloc.h> |
michael@0 | 51 | #define alloca _alloca |
michael@0 | 52 | #endif |
michael@0 | 53 | |
michael@0 | 54 | using namespace soundtouch; |
michael@0 | 55 | |
michael@0 | 56 | /***************************************************************************** |
michael@0 | 57 | * |
michael@0 | 58 | * Implementation of the class 'FIRFilter' |
michael@0 | 59 | * |
michael@0 | 60 | *****************************************************************************/ |
michael@0 | 61 | |
michael@0 | 62 | FIRFilter::FIRFilter() |
michael@0 | 63 | { |
michael@0 | 64 | resultDivFactor = 0; |
michael@0 | 65 | resultDivider = 0; |
michael@0 | 66 | length = 0; |
michael@0 | 67 | lengthDiv8 = 0; |
michael@0 | 68 | filterCoeffs = NULL; |
michael@0 | 69 | } |
michael@0 | 70 | |
michael@0 | 71 | |
michael@0 | 72 | FIRFilter::~FIRFilter() |
michael@0 | 73 | { |
michael@0 | 74 | delete[] filterCoeffs; |
michael@0 | 75 | } |
michael@0 | 76 | |
michael@0 | 77 | // Usual C-version of the filter routine for stereo sound |
michael@0 | 78 | uint FIRFilter::evaluateFilterStereo(SAMPLETYPE *dest, const SAMPLETYPE *src, uint numSamples) const |
michael@0 | 79 | { |
michael@0 | 80 | uint i, j, end; |
michael@0 | 81 | LONG_SAMPLETYPE suml, sumr; |
michael@0 | 82 | #ifdef SOUNDTOUCH_FLOAT_SAMPLES |
michael@0 | 83 | // when using floating point samples, use a scaler instead of a divider |
michael@0 | 84 | // because division is much slower operation than multiplying. |
michael@0 | 85 | double dScaler = 1.0 / (double)resultDivider; |
michael@0 | 86 | #endif |
michael@0 | 87 | |
michael@0 | 88 | assert(length != 0); |
michael@0 | 89 | assert(src != NULL); |
michael@0 | 90 | assert(dest != NULL); |
michael@0 | 91 | assert(filterCoeffs != NULL); |
michael@0 | 92 | |
michael@0 | 93 | end = 2 * (numSamples - length); |
michael@0 | 94 | |
michael@0 | 95 | for (j = 0; j < end; j += 2) |
michael@0 | 96 | { |
michael@0 | 97 | const SAMPLETYPE *ptr; |
michael@0 | 98 | |
michael@0 | 99 | suml = sumr = 0; |
michael@0 | 100 | ptr = src + j; |
michael@0 | 101 | |
michael@0 | 102 | for (i = 0; i < length; i += 4) |
michael@0 | 103 | { |
michael@0 | 104 | // loop is unrolled by factor of 4 here for efficiency |
michael@0 | 105 | suml += ptr[2 * i + 0] * filterCoeffs[i + 0] + |
michael@0 | 106 | ptr[2 * i + 2] * filterCoeffs[i + 1] + |
michael@0 | 107 | ptr[2 * i + 4] * filterCoeffs[i + 2] + |
michael@0 | 108 | ptr[2 * i + 6] * filterCoeffs[i + 3]; |
michael@0 | 109 | sumr += ptr[2 * i + 1] * filterCoeffs[i + 0] + |
michael@0 | 110 | ptr[2 * i + 3] * filterCoeffs[i + 1] + |
michael@0 | 111 | ptr[2 * i + 5] * filterCoeffs[i + 2] + |
michael@0 | 112 | ptr[2 * i + 7] * filterCoeffs[i + 3]; |
michael@0 | 113 | } |
michael@0 | 114 | |
michael@0 | 115 | #ifdef SOUNDTOUCH_INTEGER_SAMPLES |
michael@0 | 116 | suml >>= resultDivFactor; |
michael@0 | 117 | sumr >>= resultDivFactor; |
michael@0 | 118 | // saturate to 16 bit integer limits |
michael@0 | 119 | suml = (suml < -32768) ? -32768 : (suml > 32767) ? 32767 : suml; |
michael@0 | 120 | // saturate to 16 bit integer limits |
michael@0 | 121 | sumr = (sumr < -32768) ? -32768 : (sumr > 32767) ? 32767 : sumr; |
michael@0 | 122 | #else |
michael@0 | 123 | suml *= dScaler; |
michael@0 | 124 | sumr *= dScaler; |
michael@0 | 125 | #endif // SOUNDTOUCH_INTEGER_SAMPLES |
michael@0 | 126 | dest[j] = (SAMPLETYPE)suml; |
michael@0 | 127 | dest[j + 1] = (SAMPLETYPE)sumr; |
michael@0 | 128 | } |
michael@0 | 129 | return numSamples - length; |
michael@0 | 130 | } |
michael@0 | 131 | |
michael@0 | 132 | |
michael@0 | 133 | |
michael@0 | 134 | |
michael@0 | 135 | // Usual C-version of the filter routine for mono sound |
michael@0 | 136 | uint FIRFilter::evaluateFilterMono(SAMPLETYPE *dest, const SAMPLETYPE *src, uint numSamples) const |
michael@0 | 137 | { |
michael@0 | 138 | uint i, j, end; |
michael@0 | 139 | LONG_SAMPLETYPE sum; |
michael@0 | 140 | #ifdef SOUNDTOUCH_FLOAT_SAMPLES |
michael@0 | 141 | // when using floating point samples, use a scaler instead of a divider |
michael@0 | 142 | // because division is much slower operation than multiplying. |
michael@0 | 143 | double dScaler = 1.0 / (double)resultDivider; |
michael@0 | 144 | #endif |
michael@0 | 145 | |
michael@0 | 146 | |
michael@0 | 147 | assert(length != 0); |
michael@0 | 148 | |
michael@0 | 149 | end = numSamples - length; |
michael@0 | 150 | for (j = 0; j < end; j ++) |
michael@0 | 151 | { |
michael@0 | 152 | sum = 0; |
michael@0 | 153 | for (i = 0; i < length; i += 4) |
michael@0 | 154 | { |
michael@0 | 155 | // loop is unrolled by factor of 4 here for efficiency |
michael@0 | 156 | sum += src[i + 0] * filterCoeffs[i + 0] + |
michael@0 | 157 | src[i + 1] * filterCoeffs[i + 1] + |
michael@0 | 158 | src[i + 2] * filterCoeffs[i + 2] + |
michael@0 | 159 | src[i + 3] * filterCoeffs[i + 3]; |
michael@0 | 160 | } |
michael@0 | 161 | #ifdef SOUNDTOUCH_INTEGER_SAMPLES |
michael@0 | 162 | sum >>= resultDivFactor; |
michael@0 | 163 | // saturate to 16 bit integer limits |
michael@0 | 164 | sum = (sum < -32768) ? -32768 : (sum > 32767) ? 32767 : sum; |
michael@0 | 165 | #else |
michael@0 | 166 | sum *= dScaler; |
michael@0 | 167 | #endif // SOUNDTOUCH_INTEGER_SAMPLES |
michael@0 | 168 | dest[j] = (SAMPLETYPE)sum; |
michael@0 | 169 | src ++; |
michael@0 | 170 | } |
michael@0 | 171 | return end; |
michael@0 | 172 | } |
michael@0 | 173 | |
michael@0 | 174 | |
michael@0 | 175 | uint FIRFilter::evaluateFilterMulti(SAMPLETYPE *dest, const SAMPLETYPE *src, uint numSamples, uint numChannels) const |
michael@0 | 176 | { |
michael@0 | 177 | uint i, j, end, c; |
michael@0 | 178 | LONG_SAMPLETYPE *sum=(LONG_SAMPLETYPE*)alloca(numChannels*sizeof(*sum)); |
michael@0 | 179 | #ifdef SOUNDTOUCH_FLOAT_SAMPLES |
michael@0 | 180 | // when using floating point samples, use a scaler instead of a divider |
michael@0 | 181 | // because division is much slower operation than multiplying. |
michael@0 | 182 | double dScaler = 1.0 / (double)resultDivider; |
michael@0 | 183 | #endif |
michael@0 | 184 | |
michael@0 | 185 | assert(length != 0); |
michael@0 | 186 | assert(src != NULL); |
michael@0 | 187 | assert(dest != NULL); |
michael@0 | 188 | assert(filterCoeffs != NULL); |
michael@0 | 189 | |
michael@0 | 190 | end = numChannels * (numSamples - length); |
michael@0 | 191 | |
michael@0 | 192 | for (c = 0; c < numChannels; c ++) |
michael@0 | 193 | { |
michael@0 | 194 | sum[c] = 0; |
michael@0 | 195 | } |
michael@0 | 196 | |
michael@0 | 197 | for (j = 0; j < end; j += numChannels) |
michael@0 | 198 | { |
michael@0 | 199 | const SAMPLETYPE *ptr; |
michael@0 | 200 | |
michael@0 | 201 | ptr = src + j; |
michael@0 | 202 | |
michael@0 | 203 | for (i = 0; i < length; i ++) |
michael@0 | 204 | { |
michael@0 | 205 | SAMPLETYPE coef=filterCoeffs[i]; |
michael@0 | 206 | for (c = 0; c < numChannels; c ++) |
michael@0 | 207 | { |
michael@0 | 208 | sum[c] += ptr[0] * coef; |
michael@0 | 209 | ptr ++; |
michael@0 | 210 | } |
michael@0 | 211 | } |
michael@0 | 212 | |
michael@0 | 213 | for (c = 0; c < numChannels; c ++) |
michael@0 | 214 | { |
michael@0 | 215 | #ifdef SOUNDTOUCH_INTEGER_SAMPLES |
michael@0 | 216 | sum[c] >>= resultDivFactor; |
michael@0 | 217 | #else |
michael@0 | 218 | sum[c] *= dScaler; |
michael@0 | 219 | #endif // SOUNDTOUCH_INTEGER_SAMPLES |
michael@0 | 220 | *dest = (SAMPLETYPE)sum[c]; |
michael@0 | 221 | dest++; |
michael@0 | 222 | sum[c] = 0; |
michael@0 | 223 | } |
michael@0 | 224 | } |
michael@0 | 225 | return numSamples - length; |
michael@0 | 226 | } |
michael@0 | 227 | |
michael@0 | 228 | |
michael@0 | 229 | // Set filter coeffiecients and length. |
michael@0 | 230 | // |
michael@0 | 231 | // Throws an exception if filter length isn't divisible by 8 |
michael@0 | 232 | void FIRFilter::setCoefficients(const SAMPLETYPE *coeffs, uint newLength, uint uResultDivFactor) |
michael@0 | 233 | { |
michael@0 | 234 | assert(newLength > 0); |
michael@0 | 235 | if (newLength % 8) ST_THROW_RT_ERROR("FIR filter length not divisible by 8"); |
michael@0 | 236 | |
michael@0 | 237 | lengthDiv8 = newLength / 8; |
michael@0 | 238 | length = lengthDiv8 * 8; |
michael@0 | 239 | assert(length == newLength); |
michael@0 | 240 | |
michael@0 | 241 | resultDivFactor = uResultDivFactor; |
michael@0 | 242 | resultDivider = (SAMPLETYPE)::pow(2.0, (int)resultDivFactor); |
michael@0 | 243 | |
michael@0 | 244 | delete[] filterCoeffs; |
michael@0 | 245 | filterCoeffs = new SAMPLETYPE[length]; |
michael@0 | 246 | memcpy(filterCoeffs, coeffs, length * sizeof(SAMPLETYPE)); |
michael@0 | 247 | } |
michael@0 | 248 | |
michael@0 | 249 | |
michael@0 | 250 | uint FIRFilter::getLength() const |
michael@0 | 251 | { |
michael@0 | 252 | return length; |
michael@0 | 253 | } |
michael@0 | 254 | |
michael@0 | 255 | |
michael@0 | 256 | |
michael@0 | 257 | // Applies the filter to the given sequence of samples. |
michael@0 | 258 | // |
michael@0 | 259 | // Note : The amount of outputted samples is by value of 'filter_length' |
michael@0 | 260 | // smaller than the amount of input samples. |
michael@0 | 261 | uint FIRFilter::evaluate(SAMPLETYPE *dest, const SAMPLETYPE *src, uint numSamples, uint numChannels) const |
michael@0 | 262 | { |
michael@0 | 263 | assert(length > 0); |
michael@0 | 264 | assert(lengthDiv8 * 8 == length); |
michael@0 | 265 | |
michael@0 | 266 | if (numSamples < length) return 0; |
michael@0 | 267 | |
michael@0 | 268 | #ifndef USE_MULTICH_ALWAYS |
michael@0 | 269 | if (numChannels == 1) |
michael@0 | 270 | { |
michael@0 | 271 | return evaluateFilterMono(dest, src, numSamples); |
michael@0 | 272 | } |
michael@0 | 273 | else if (numChannels == 2) |
michael@0 | 274 | { |
michael@0 | 275 | return evaluateFilterStereo(dest, src, numSamples); |
michael@0 | 276 | } |
michael@0 | 277 | else |
michael@0 | 278 | #endif // USE_MULTICH_ALWAYS |
michael@0 | 279 | { |
michael@0 | 280 | assert(numChannels > 0); |
michael@0 | 281 | return evaluateFilterMulti(dest, src, numSamples, numChannels); |
michael@0 | 282 | } |
michael@0 | 283 | } |
michael@0 | 284 | |
michael@0 | 285 | |
michael@0 | 286 | |
michael@0 | 287 | // Operator 'new' is overloaded so that it automatically creates a suitable instance |
michael@0 | 288 | // depending on if we've a MMX-capable CPU available or not. |
michael@0 | 289 | void * FIRFilter::operator new(size_t s) |
michael@0 | 290 | { |
michael@0 | 291 | // Notice! don't use "new FIRFilter" directly, use "newInstance" to create a new instance instead! |
michael@0 | 292 | ST_THROW_RT_ERROR("Error in FIRFilter::new: Don't use 'new FIRFilter', use 'newInstance' member instead!"); |
michael@0 | 293 | return newInstance(); |
michael@0 | 294 | } |
michael@0 | 295 | |
michael@0 | 296 | |
michael@0 | 297 | FIRFilter * FIRFilter::newInstance() |
michael@0 | 298 | { |
michael@0 | 299 | #if defined(SOUNDTOUCH_ALLOW_MMX) || defined(SOUNDTOUCH_ALLOW_SSE) |
michael@0 | 300 | uint uExtensions; |
michael@0 | 301 | |
michael@0 | 302 | uExtensions = detectCPUextensions(); |
michael@0 | 303 | #endif |
michael@0 | 304 | |
michael@0 | 305 | // Check if MMX/SSE instruction set extensions supported by CPU |
michael@0 | 306 | |
michael@0 | 307 | #ifdef SOUNDTOUCH_ALLOW_MMX |
michael@0 | 308 | // MMX routines available only with integer sample types |
michael@0 | 309 | if (uExtensions & SUPPORT_MMX) |
michael@0 | 310 | { |
michael@0 | 311 | return ::new FIRFilterMMX; |
michael@0 | 312 | } |
michael@0 | 313 | else |
michael@0 | 314 | #endif // SOUNDTOUCH_ALLOW_MMX |
michael@0 | 315 | |
michael@0 | 316 | #ifdef SOUNDTOUCH_ALLOW_SSE |
michael@0 | 317 | if (uExtensions & SUPPORT_SSE) |
michael@0 | 318 | { |
michael@0 | 319 | // SSE support |
michael@0 | 320 | return ::new FIRFilterSSE; |
michael@0 | 321 | } |
michael@0 | 322 | else |
michael@0 | 323 | #endif // SOUNDTOUCH_ALLOW_SSE |
michael@0 | 324 | |
michael@0 | 325 | { |
michael@0 | 326 | // ISA optimizations not supported, use plain C version |
michael@0 | 327 | return ::new FIRFilter; |
michael@0 | 328 | } |
michael@0 | 329 | } |